CoolFace
Modelpublic

SZLHOLDINGS/SZL-Khipu-1.5B-abstain

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face
1likes20downloads
Model Card
EXPERIMENT. Adapter bytes missing or unverified. Evaluators use SZLHOLDINGS/SZL-Khipu-1.5B until a receipted adapter exists.
NO WEIGHTS IN THIS REPO — metadata corrected. The card already said "WEIGHTS UNAVAILABLE", but the front matter simultaneously declared library_name: peft, base_model_relation: adapter and pipeline_tag: text-generation, plus peft/qlora tags. Together those tell the Hub this is a loadable PEFT adapter. There is no adapter_model.safetensors here, so it is not. Those four declarations have been removed; the prose was already honest and is unchanged. What is here is a complete, runnable training curriculum: train.jsonl, train.abstain.jsonl, adversarial.jsonl, eval.jsonl, a 23 KB training script, and a manifest. Everything needed to produce the adapter is present — it has simply not been run. The trained successor is KHIPU-R2 (abstain 3/6 MEASURED, declared not a pass).

SZL-Khipu-1.5B-abstain

WEIGHTS UNAVAILABLE. No adapter_model.safetensors on this ID. Successor adapter with weights is `SZLHOLDINGS/KHIPU-R2`.

QLoRA adapter retrain recipe of the existing Khipu line. Raises in-memory ABSTAIN_OVERSAMPLE from 2 to 4 (8×4=32 abstain vs 15 navigate). Proposal-only. Λ = Conjecture 1. Doctrine v11 LOCKED 749/14/163.

This ID currently holds curriculum + script only. It is not a loadable PEFT adapter.

WeightsUNAVAILABLE

<!-- SZL-ATELIER-CUT:v1:START -->

The cut

Most 'safety LoRAs' teach tone. This one teaches a binary: the handles are not enough. Research-only until abstain beats 2/6.

A specialist in silence. Capability is someone else's LoRA.

Silhouette → leave → SZL

LeaderTake, then tweak
AnthropicConstitutional fine-tune, but only the refuse clause.
NVIDIAA guardrail as weights, not as Colang.
UnslothQLoRA adapter, proposal-only, research-only tag.

Nobody else ships this combination. That is the point of a one-of-one.

Intended use

Stack on the navigator. Measure abstain. Do not ship on hope.

Limitations

  • —research-only
  • —proposal-only
  • —Does not magically fix 2/6 until a signed eval says so.

Canonical GitHub: `szl-holdings/szl-forge` <!-- SZL-ATELIER-CUT:v1:END -->

| Jobs | UNAVAILABLE | | Base (canonical) | `Qwen/Qwen2.5-1.5B-Instruct` | | Runtime train | unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit (same Qwen2.5-1.5B weights, 4-bit) | | Relation | adapter (declared; files not present) | | License | Apache-2.0 | | Does NOT overwrite | `SZLHOLDINGS/SZL-Khipu-1.5B` signed weights | | Successor with weights | `SZLHOLDINGS/KHIPU-R2` (MEASURED abstain 3/6, not a pass) | | This is NOT | the Chaski Qwen3.5 lock |

Evaluation

Status: NOT YET RUN. No fabricated k/n. publication_eligible is false until the held-out eval in train_khipu_abstain.py actually executes after training.

Prior original MEASURED abstain on SZLHOLDINGS/SZL-Khipu-1.5B is 2/6 (blocker). Eval protocol: eval.jsonl 5 navigate + adversarial.jsonl 6 abstain. Report k/n only.

Training (this job)

  • —Unsloth QLoRA, seed 11, lr 2e-4, adamw8bit, `trainonresponsesonly`, Trackio
  • —LoRA r=32 α=64, 45 epochs, ga=2, batch=1, constantwithwarmup (from train_khipu.py)
  • —Train: train.jsonl 15 navigate + train.abstain.jsonl 8 rows × 4
  • —Held-out never in gradients
  • —Script: `train_khipu_abstain.py`

Intended use

Proposal-only JSON retrieval plans (NAVIGATE / ABSTAIN) over synthetic Brain node handles. A controller outside the weights validates and resolves content. Not autonomous. Not a replacement for the signed original weights.